[Difx-users] Speedup Factor plot_logtime.py

[Adam Deller]

Depends on what people want to use the reported speedup for, I suppose.
The easy way is to simply report sum(obs_time) / max(correlate_time), like
you suggest.  That naturally takes into account inefficiencies in
allocation across multiple jobs and gives a 'real world' speedup.  So if
someone splits their jobs in such a way that one job takes much longer,
then the speedup reported is representative of the real-world time taken to
finish everything.  I think this is a good default.

That's one extreme.  The other extreme is if you assume that as soon as one
job finishes someone else takes those resources and starts using them for
something else useful, and so you want to report the weighted average
speedup of the individual jobs.  Each individual job has a speedup of s_i =
obs_time_i/correlate_time_i, and then you form the overall speedup by
sum(s_i * obs_time_i) / sum(obs_time_i).

Those two measures are useful for different things, so probably it would be
good to report both.  But I think for most people the first is more useful.

Cheers,
Adam

On 28 April 2017 at 08:37, Stuart Weston <stuart.weston at aut.ac.nz> wrote:

> The simple one head -> many workers we have speedup factor:
>
>
>
> Speedup = obs_time / corr_time
>
>
>
> So say 4hrs observing is correlated in ~10.3 hrs, a speedup of 0.388.
>
>
>
> Now in the scenario where we split the job between multiple heads, so
> head_1 does 0-2 hrs and head_2 does 2-4 hrs. How do we want to define the
> speedup here ?
>
>
>
> Each will take slightly different times due to many external factors, but
> both are started at the same time (within fractions of a second)
>
>
>
> So head_1 takes 5.1 hrs to correlate 2 hrs of observing, and head_2 takes
> 5.3 hrs for the other 2 hrs.
>
>
>
> We have correlated the same 4hrs of data but in approx half the elapsed
> time. So would you be happy for me to define speedup now as:
>
>
>
> Speedup = sum(obs_time_1, obs_time_2) / max(corr_time(head_1,head_2))
>
>
>
> In this case it would be say 4/5.3 = 0.75. Now this doesn’t work if the
> split is un-even for the obs_time between the two heads.
>
>
>
> Just looking for a way to define my benchmarks (agreeable to the
> community) for cloud correlation and using multiple heads and breaking a
> job up between heads on different hypervisors.
>
>
>
> Stuart Weston Bsc (Hons), MPhil (Hons), MInstP
>
> Mobile: 021 713062
>
> Skype: stuart.d.weston
> Email:  stuart.weston at aut.ac.nz
>
> http://www.atnf.csiro.au/people/Stuart.Weston/index.html
>
> Software Engineer
> Institute for Radio Astronomy & Space Research (IRASR)
> School of Computing & Mathematical Sciences
> Faculty of Creative Technologies
> Auckland University of Technology, New Zealand.
>
> http://www.irasr.aut.ac.nz/
>
>
>
> [image: NewIRASRLogo]
>
>
>
>
>
> _______________________________________________
> Difx-users mailing list
> Difx-users at listmgr.nrao.edu
> https://listmgr.nrao.edu/mailman/listinfo/difx-users
>
>


-- 
!=============================================================!
Dr. Adam Deller
ARC Future Fellow, Senior Lecturer
Centre for Astrophysics & Supercomputing
Swinburne University of Technology
John St, Hawthorn VIC 3122 Australia
phone: +61 3 9214 5307
fax: +61 3 9214 8797

office days (usually): Mon-Thu
!=============================================================!
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listmgr.nrao.edu/pipermail/difx-users/attachments/20170428/2be342e5/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image003.jpg
Type: image/jpeg
Size: 11334 bytes
Desc: not available
URL: <http://listmgr.nrao.edu/pipermail/difx-users/attachments/20170428/2be342e5/attachment-0001.jpg>